Pigmented Dermatosis Recognition System Based On Semi-supervision

In the last decades,the number of dermatosis patients has been growing rapidly all over the world.But the corresponding medical facilities and awareness have not improved at the same time,and patients generally cannot receive adequate treatment.In the same time,the research results of the use of artificial intelligence algorithms for auxiliary diagnosis and treatment are increasingly outstanding,but the ability to directly provide auxiliary diagnosis services is generally lacking.Therefore,the purpose of this paper is to realize an pigmented dermatosis recognition system to provide disease monitoring services by itself,On the other hand,the medical field also has problems of high cost and difficulty in labeling data,which also makes the research to solve a small number of datasets.All these factors makes it necessary to research how to solve the question of lacking of data to train.The main research of this paper includes used the CNN to identify pigmented dermatosis,used GAN to realize the semi-supervised mechanism to solve the problem of lacking train data,and built a complete pigmented dermatosis recognition system based on wechat mini program and a corresponding management system based on web.So it can be used to assist recognition the diagnosis of four pigmented dermatosis including melanoma,vitiligo,melanocytic nevus and pigmented benign keratosis.The innovation points involved are summarized as the following three points:（1）This paper designed a GAN which separated discriminator network into identification and classification networks to achieve semi-supervision mechanism and used larger convolution kernels and expand network level to improve the learning ability of the network itself,so the model can get the effect of higher accuracy and precision on the basis of using just about 1500 labeled data to train.（2）The paper combined the attention mechanism and multi convolutional channel mechanism to modified the original classifier network structure so that it can use multi-branch-based attention network helps the model obtain more abundant image feature information.（3）Based on the above algorithm research,this paper designed and realize a semi-supervised pigmented dermatosis recognition system and successfully deployed it to realize real-time auxiliary diagnosis and realize the engineering implementation of the algorithm.The pigmented dermatosis recognition system realized in this paper can achieve the recall rate,precision and accuracy rate are more than 91% by using just 4% data of dataset Skin Cancer MNIST: HAM10000 and 1/4 data of Skin Cancer ISIC,In addition to support convenient diagnosis and detection services,it can also provide related information services and treatment suggestions based on the mini program.The system has great medical value and social value.